Meta has introduced Muse Wonder and Muse Code, models designed for agentic tasks that can extend across documents, tools and time. The company says developers can access the models through its API and use Muse Code for software work, while examples include insurance claims, lease analysis and financial tasks.
The product concept relies on “Active Memory” and “Active Focus”, according to Meta. That makes context management a governance issue, not merely a convenience feature. Persistent memory can improve continuity, but it also creates questions about data provenance, retention, deletion, access rights and the possibility that an incorrect or sensitive detail shapes later actions.
Meta says the models were trained to avoid sharing private information and causing harm. It also says a safety report will be available soon. Until that report is published, customers should treat the release note as a description of intended safeguards rather than a complete assurance package.
The enterprise control checklist is straightforward: define what the agent may remember, separate short-lived task context from durable memory, give users a way to inspect and correct stored information, and test whether deletion propagates through indexes and downstream tools. Long-running agents need memory governance that is as explicit as their tool permissions.
